Selection-Based Image Generation for Semantic Communication Systems

Efficient image transmission while preserving semantic information is crucial for many applications but poses challenges when communication channels have limited capacity. This letter presents an end-to-end semantic communication system for image transmission by exploiting semantic information, which is different from traditional approaches that emphasize pixel-level information. Specifically, identical semantic knowledge libraries are shared between the sender and receiver to associate image contents with semantics. At the sender, a deep learning-based classifier categorizes the image, and a dictionary learning method extracts features. At the receiver, a modified diffusion model-based generator reconstructs the image from the received features and category, with the objective to minimize the reconstruction error. To evaluate the semantic fidelity, we propose a semantic fidelity index (SFI) that considers both mutual information and neural network (NN) feature similarity between the original and reconstructed images. Experiments demonstrate that, by leveraging the shared semantic prior knowledge base, our approach can efficiently convey image semantics and achieve high-quality reconstruction. The proposed system provides an effective solution for semantic-preserving image communication in bandwidth-limited applications.